DOCSLIB.ORG

Learning to Rank Learning Curves

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Learning to Rank Learning Curves Learning to Rank Learning Curves Martin Wistuba 1 Tejaswini Pedapati 1 Abstract predictions (Chandrashekaran & Lane, 2017; Klein et al., 2017; Baker et al., 2018). Thus, the extrapolation method Many automated machine learning methods, such for the first candidates can not be used yet, which means as those for hyperparameter and neural archi- more computational effort. Other methods do not have this tecture optimization, are computationally expen- disadvantage, but require sufficiently long learning curves sive because they involve training many different to make reliable predictions which again means unnecessary model configurations. In this work, we present a overhead (Domhan et al., 2015). Many of these methods new method that saves computational budget by also do not take into account other information such as the terminating poor configurations early on in the hyperparameters of the model being examined or its network training. In contrast to existing methods, we con- architecture. sider this task as a ranking and transfer learning problem. We qualitatively show that by optimiz- We address the need for sample learning curves by devising ing a pairwise ranking loss and leveraging learn- a transfer learning technique that uses learning curves from ing curves from other datasets, our model is able other problems. Since the range of accuracy varies from to effectively rank learning curves without hav- dataset to dataset, we are forced to consider this in our ing to observe many or very long learning curves. modeling. But since we are not interested in predicting the We further demonstrate that our method can be performance of a model anyway, we use a ranking model used to accelerate a neural architecture search by that models the probability that the model currently being a factor of up to 100 without a significant perfor- investigated surpasses the best solution so far. mance degradation of the discovered architecture. In order to be able to make reliable predictions for short In further experiments we analyze the quality of learning curves, we consider further characteristics of the ranking, the influence of different model compo- model such as its network architecture. We compare our nents as well as the predictive behavior of the ranking method with respect to a ranking measure against model. different methods on five different image classification and four tabular regression datasets. We also show that our method is capable of significantly accelerating neural ar- 1. Introduction chitecture search (NAS) and hyperparameter optimization. A method commonly used by human experts to speed up Furthermore, we conduct several ablation studies to provide the optimization of neural architectures or hyperparameters a better motivation of our model and its behavior. is the early termination of iterative training processes that are unlikely to improve the current solution. A common 2. Related work arXiv:2006.03361v1 [cs.LG] 5 Jun 2020 technique to determine the likelihood of no improvement is to compare the learning curve of a new configuration to the Most of the prior work for learning curve prediction is based one of the currently best configuration. This idea can also be on the idea of extrapolating the partial learning curve by used to speed up automated machine learning processes. For using a combination of continuously increasing basic func- this purpose, it is common practice to extrapolate the partial tions. learning curve in order to predict the final performance of Domhan et al.(2015) define a set of 11 parametric basic the currently investigated model. Current extrapolation tech- functions, estimate their parameters and combine them in niques have several weaknesses that make them unable to an ensemble. Klein et al.(2017) propose a heteroscedastic realize their full potential in practice. Many of the methods Bayesian model which learns a weighted average of the require sufficient sample learning curves to make reliable basic functions. Chandrashekaran & Lane(2017) do not 1IBM Research. Correspondence to: Martin Wistuba <mar- use basic functions but use previously observed learning [email protected]>. curves of the current dataset. An affine transformation for each previously seen learning curve is estimated by mini- Learning to Rank Learning Curves mizing the mean squared error with respect to the partial × learning curve. The best fitting extrapolations are averaged Final Performance × as the final prediction. Baker et al.(2018) use a different Intermediate Performance procedure. They use support vector machines as sequential regressive models to predict the final accuracy based on features extracted from the learning curves, its gradients, and the neural architecture itself. Validation Accuracy Iterations The predictor by Domhan et al.(2015) is able to forecast without seeing any learning curve before but requires ob- serving more epochs for accurate predictions. The model by Figure 1. Learning curve prediction tries to predict from the partial Chandrashekaran & Lane(2017) requires seeing few learn- learning curve (solid line) the final performance. ing curves to extrapolate future learning curves. However, accurate forecasts are already possible after few epochs. Al- gorithms proposed by Klein et al.(2017); Baker et al.(2018) need to observe many full-length learning curves before pro- viding any useful forecasts. However, this is prohibiting in learned and finally used to determine the best architecture. the scenarios where learning is time-consuming such as in DARTS (Liu et al., 2019) is one of the last extensions of large convolutional neural networks (CNN). this idea which uses a continuous relaxation of the search All previous methods for automatically terminating iterative space, which allows learning of the shared parameters as learning processes are based on methods that predict the well as of the architecture by means of gradient descent. learning curve. Ultimately, however, we are less interested These methods can be a strong alternative to early termi- in the exact learning curve but rather whether the current nation but do not transfer to general machine learning and learning curve leads to a better result. This way of obtaining recent research implies that this approach does not work for a ranking is referred to as pointwise ranking methods (Liu, arbitrary architecture search spaces (Yu et al., 2020). 2011). They have proven to be less efficient than pairwise ranking methods which directly optimize for the objective 3. Learning curve ranking function (Burges et al., 2005). Yet, we are the first to con- sider a pairwise ranking loss for this application. With learning curve we refer to the function of qualitative performance with growing number of iterations of an itera- There are some bandit-based methods which leverage early tive learning algorithm. We use the term final learning curve termination as well. Successive Halving (Jamieson & Tal- to explicitly denote the entire learning curve, y1; : : : ; yL, re- walkar, 2016) is a method that trains multiple models with flecting the training process from beginning to end. Here, yi settings chosen at random simultaneously and terminates is a measure of the performance of the model (e.g., classifi- the worst performing half in predefined intervals. Hyper- cation accuracy), which is determined at regular intervals. band (Li et al., 2017) identifies that the choice of the inter- On the contrary, a partial learning curve, y1; : : : ; yl, refers vals is vital and therefore proposes to run Successive Halv- to learning curves that are observed only up to a time l. We ing with different intervals. BOHB (Falkner et al., 2018) is visualize the concepts of the terms in Figure1. an extension of Hyperband which proposes to replace the random selection of settings with Bayesian optimization. There is a set of automated early termination methods which follow broadly the same principle. The first model is trained The use of methods that terminate less promising iterative to completion and is considered as the current best model training processes early are of particular interest in the field mmax with its performance being ymax. For each further of automated machine learning, as they can, for example, model mi, the probability that it is better than the current significantly accelerate hyperparameter optimization. The max best model, p (mi > m ), is monitored at periodic inter- most computationally intensive subproblem of automated vals during training. If it is below a given threshold, the machine learning is Neural Architecture Search (Wistuba model’s training is terminated (see Algorithm1). All exist- et al., 2019), the optimization of the neural network topol- ing methods rely on a two-step approach to determine this ogy. Our method is not limited to this problem, but as it is probability which involves extrapolating the partial learning currently one of the major challenges in automated machine curve and a heuristic measure. Instead of the two-step pro- learning, we use this problem as a sample application in cess to determine the probability, we propose LCRankNet our evaluation. For this particular application, optimization to predict the probability that a model mi is better than methods that leverage parameter sharing (Pham et al., 2018) mj directly. LCRankNet is based on a neural network f have become established as a standard method. Here, a which considers model characteristics and the partial learn- search space spanning, overparameterized neural network is ing curve xi as input. We define the probability that mi is Learning to Rank Learning Curves better than mj as f(x) ef(xi)−f(xj ) p(mi > mj) =p ^i;j = : (1) Fully Connected Layer 1 + ef(xi)−f(xj ) Using the logistic function is an established modelling ap- max proach in the learning-to-rank community (Burges et al., LSTM pool 2005). Given a set of final learning curves for some models, conv the estimation of the values p between these models is Embedding Embedding i;j k=2..5 trivial since we know whether mi is better than mj or not.
Load more

Recommended publications
  • Context-Aware Learning to Rank with Self-Attention
    Context-Aware Learning to Rank with Self-Attention Przemysław Pobrotyn Tomasz Bartczak Mikołaj Synowiec ML Research at Allegro.pl ML Research at Allegro.pl ML Research at Allegro.pl [email protected] [email protected] [email protected] Radosław Białobrzeski Jarosław Bojar ML Research at Allegro.pl ML Research at Allegro.pl [email protected] [email protected] ABSTRACT engines, recommender systems, question-answering systems, and Learning to rank is a key component of many e-commerce search others. engines. In learning to rank, one is interested in optimising the A typical machine learning solution to the LTR problem involves global ordering of a list of items according to their utility for users. learning a scoring function, which assigns real-valued scores to Popular approaches learn a scoring function that scores items in- each item of a given list, based on a dataset of item features and dividually (i.e. without the context of other items in the list) by human-curated or implicit (e.g. clickthrough logs) relevance labels. optimising a pointwise, pairwise or listwise loss. The list is then Items are then sorted in the descending order of scores [23]. Perfor- sorted in the descending order of the scores. Possible interactions mance of the trained scoring function is usually evaluated using an between items present in the same list are taken into account in the IR metric like Mean Reciprocal Rank (MRR) [33], Normalised Dis- training phase at the loss level. However, during inference, items counted Cumulative Gain (NDCG) [19] or Mean Average Precision are scored individually, and possible interactions between them are (MAP) [4].
    [Show full text]
  • An Introduction to Neural Information Retrieval
    The version of record is available at: http://dx.doi.org/10.1561/1500000061 An Introduction to Neural Information Retrieval Bhaskar Mitra1 and Nick Craswell2 1Microsoft, University College London; [email protected] 2Microsoft; [email protected] ABSTRACT Neural ranking models for information retrieval (IR) use shallow or deep neural networks to rank search results in response to a query. Tradi- tional learning to rank models employ supervised machine learning (ML) techniques—including neural networks—over hand-crafted IR features. By contrast, more recently proposed neural models learn representations of language from raw text that can bridge the gap between query and docu- ment vocabulary. Unlike classical learning to rank models and non-neural approaches to IR, these new ML techniques are data-hungry, requiring large scale training data before they can be deployed. This tutorial intro- duces basic concepts and intuitions behind neural IR models, and places them in the context of classical non-neural approaches to IR. We begin by introducing fundamental concepts of retrieval and different neural and non-neural approaches to unsupervised learning of vector representations of text. We then review IR methods that employ these pre-trained neural vector representations without learning the IR task end-to-end. We intro- duce the Learning to Rank (LTR) framework next, discussing standard loss functions for ranking. We follow that with an overview of deep neural networks (DNNs), including standard architectures and implementations. Finally, we review supervised neural learning to rank models, including re- cent DNN architectures trained end-to-end for ranking tasks. We conclude with a discussion on potential future directions for neural IR.
    [Show full text]
  • Metric Learning for Session-Based Recommendations - Preprint
    METRIC LEARNING FOR SESSION-BASED RECOMMENDATIONS A PREPRINT Bartłomiej Twardowski1,2 Paweł Zawistowski Szymon Zaborowski 1Computer Vision Center, Warsaw University of Technology, Sales Intelligence Universitat Autónoma de Barcelona Institute of Computer Science 2Warsaw University of Technology, [email protected] Institute of Computer Science [email protected] January 8, 2021 ABSTRACT Session-based recommenders, used for making predictions out of users’ uninterrupted sequences of actions, are attractive for many applications. Here, for this task we propose using metric learning, where a common embedding space for sessions and items is created, and distance measures dissimi- larity between the provided sequence of users’ events and the next action. We discuss and compare metric learning approaches to commonly used learning-to-rank methods, where some synergies ex- ist. We propose a simple architecture for problem analysis and demonstrate that neither extensively big nor deep architectures are necessary in order to outperform existing methods. The experimental results against strong baselines on four datasets are provided with an ablation study. Keywords session-based recommendations · deep metric learning · learning to rank 1 Introduction We consider the session-based recommendation problem, which is set up as follows: a user interacts with a given system (e.g., an e-commerce website) and produces a sequence of events (each described by a set of attributes). Such a continuous sequence is called a session, thus we denote sk = ek,1,ek,2,...,ek,t as the k-th session in our dataset, where ek,j is the j-th event in that session. The events are usually interactions with items (e.g., products) within the system’s domain.
    [Show full text]
  • Learning Embeddings: Efficient Algorithms and Applications
    Learning embeddings: HIÀFLHQWDOJRULWKPVDQGDSSOLFDWLRQV THÈSE NO 8348 (2018) PRÉSENTÉE LE 15 FÉVRIER 2018 À LA FACULTÉ DES SCIENCES ET TECHNIQUES DE L'INGÉNIEUR LABORATOIRE DE L'IDIAP PROGRAMME DOCTORAL EN GÉNIE ÉLECTRIQUE ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE POUR L'OBTENTION DU GRADE DE DOCTEUR ÈS SCIENCES PAR Cijo JOSE acceptée sur proposition du jury: Prof. P. Frossard, président du jury Dr F. Fleuret, directeur de thèse Dr O. Bousquet, rapporteur Dr M. Cisse, rapporteur Dr M. Salzmann, rapporteur Suisse 2018 Abstract Learning to embed data into a space where similar points are together and dissimilar points are far apart is a challenging machine learning problem. In this dissertation we study two learning scenarios that arise in the context of learning embeddings and one scenario in efficiently estimating an empirical expectation. We present novel algorithmic solutions and demonstrate their applications on a wide range of data-sets. The first scenario deals with learning from small data with large number of classes. This setting is common in computer vision problems such as person re-identification and face verification. To address this problem we present a new algorithm called Weighted Approximate Rank Component Analysis (WARCA), which is scalable, robust, non-linear and is independent of the number of classes. We empirically demonstrate the performance of our algorithm on 9 standard person re-identification data-sets where we obtain state of the art performance in terms of accuracy as well as computational speed. The second scenario we consider is learning embeddings from sequences. When it comes to learning from sequences, recurrent neural networks have proved to be an effective algorithm.
    [Show full text]
  • An Empirical Study of Embedding Features in Learning to Rank
    An Empirical Study of Embedding Features in Learning to Rank ABSTRACT use of entities can recognize which part of the query is the most is paper explores the possibility of using neural embedding fea- informative; (2) the second aspect that we consider is the ‘type tures for enhancing the eectiveness of ad hoc document ranking of embeddings’ that can be used. We have included both word based on learning to rank models. We have extensively introduced embedding [8] and document embedding [4] models for building and investigated the eectiveness of features learnt based on word dierent features. and document embeddings to represent both queries and docu- Based on these two cross-cuing aspects, namely type of content ments. We employ several learning to rank methods for document and type of embedding, we have dened ve classes of features as ranking using embedding-based features, keyword-based features shown in Table 1. Briey stated, our features are query-dependent as well as the interpolation of the embedding-based features with features and hence, we employ the embedding models to form vec- keyword-based features. e results show that embedding features tor representations of words, chunks and entities in both documents have a synergistic impact on keyword based features and are able and queries that can then be employed for measuring the distance to provide statistically signicant improvement on harder queries. between queries and documents. It should be noted that we do not use entity embeddings [3] to represent entities, but rather use the ACM Reference format: same word embedding and document embedding models on the .
    [Show full text]
  • Learning to Rank Using Gradient Descent
    Learning to Rank using Gradient Descent Keywords: ranking, gradient descent, neural networks, probabilistic cost functions, internet search Chris Burges [email protected] Tal Shaked∗ [email protected] Erin Renshaw [email protected] Microsoft Research, One Microsoft Way, Redmond, WA 98052-6399 Ari Lazier [email protected] Matt Deeds [email protected] Nicole Hamilton [email protected] Greg Hullender [email protected] Microsoft, One Microsoft Way, Redmond, WA 98052-6399 Abstract that maps to the reals (having the model evaluate on We investigate using gradient descent meth- pairs would be prohibitively slow for many applica- ods for learning ranking functions; we pro- tions). However (Herbrich et al., 2000) cast the rank- pose a simple probabilistic cost function, and ing problem as an ordinal regression problem; rank we introduce RankNet, an implementation of boundaries play a critical role during training, as they these ideas using a neural network to model do for several other algorithms (Crammer & Singer, the underlying ranking function. We present 2002; Harrington, 2003). For our application, given test results on toy data and on data from a that item A appears higher than item B in the out- commercial internet search engine. put list, the user concludes that the system ranks A higher than, or equal to, B; no mapping to particular rank values, and no rank boundaries, are needed; to 1. Introduction cast this as an ordinal regression problem is to solve an unnecessarily hard problem, and our approach avoids Any system that presents results to a user, ordered this extra step. We also propose a natural probabilis- by a utility function that the user cares about, is per- tic cost function on pairs of examples.
    [Show full text]
  • Multi-Agent Reinforced Learning to Rank
    MarlRank: Multi-agent Reinforced Learning to Rank Shihao Zou∗ Zhonghua Li Mohammad Akbari [email protected] [email protected] [email protected] University of Alberta Huawei University College London Jun Wang Peng Zhang [email protected] [email protected] University College London Tianjin University ABSTRACT 1 INTRODUCTION When estimating the relevancy between a query and a document, Learning to rank, which learns a model for ranking documents ranking models largely neglect the mutual information among doc- (items) based on their relevance scores to a given query, is a key uments. A common wisdom is that if two documents are similar task in information retrieval (IR) [18]. Typically, a query-document in terms of the same query, they are more likely to have similar pair is represented as a feature vector and many classical models, in- relevance score. To mitigate this problem, in this paper, we propose cluding SVM [8], AdaBoost [18] and regression tree [2] are applied a multi-agent reinforced ranking model, named MarlRank. In partic- to learn to predict the relevancy. Recently, deep learning models has ular, by considering each document as an agent, we formulate the also been applied in learning to rank task. For example, CDSSM [13], ranking process as a multi-agent Markov Decision Process (MDP), DeepRank [12] and their variants aim at learning better semantic where the mutual interactions among documents are incorporated representation of the query-document pair for more accurate pre- in the ranking process. To compute the ranking list, each docu- diction of the relevancy.
    [Show full text]
  • Yahoo! Learning to Rank Challenge Overview
    JMLR: Workshop and Conference Proceedings 14 (2011) 1–24 Yahoo! Learning to Rank Challenge Yahoo! Learning to Rank Challenge Overview Olivier Chapelle⇤ [email protected] Yi Chang [email protected] Yahoo! Labs Sunnyvale, CA Abstract Learning to rank for information retrieval has gained a lot of interest in the recent years but there is a lack for large real-world datasets to benchmark algorithms. That led us to publicly release two datasets used internally at Yahoo! for learning the web search ranking function. To promote these datasets and foster the development of state-of-the-art learning to rank algorithms, we organized the Yahoo! Learning to Rank Challenge in spring 2010. This paper provides an overview and an analysis of this challenge, along with a detailed description of the released datasets. 1. Introduction Ranking is at the core of information retrieval: given a query, candidates documents have to be ranked according to their relevance to the query. Learning to rank is a relatively new field in which machine learning algorithms are used to learn this ranking function. It is of particular importance for web search engines to accurately tune their ranking functions as it directly a↵ects the search experience of millions of users. A typical setting in learning to rank is that feature vectors describing a query-document pair are constructed and relevance judgments of the documents to the query are available. A ranking function is learned based on this training data, and then applied to the test data. Several benchmark datasets, such as letor, have been released to evaluate the newly proposed learning to rank algorithms.
    [Show full text]
  • Sodeep: a Sorting Deep Net to Learn Ranking Loss Surrogates
    SoDeep: a Sorting Deep net to learn ranking loss surrogates Martin Engilberge1,2, Louis Chevallier2, Patrick Perez´ 3, Matthieu Cord1,3 1Sorbonne Universite,´ Paris, France, 2Technicolor, Cesson Sevign´ e,´ France, 3Valeo.ai, Paris, France {martin.engilberge, matthieu.cord}@lip6.fr [email protected] [email protected] Abstract Several tasks in machine learning are evaluated us- ing non-differentiable metrics such as mean average pre- cision or Spearman correlation. However, their non- differentiability prevents from using them as objective func- tions in a learning framework. Surrogate and relaxation methods exist but tend to be specific to a given metric. In the present work, we introduce a new method to learn approximations of such non-differentiable objective func- tions. Our approach is based on a deep architecture that approximates the sorting of arbitrary sets of scores. It is Figure 1: Overview of SoDeep, the proposed end-to- trained virtually for free using synthetic data. This sorting end trainable deep architecture to approximate non- deep (SoDeep) net can then be combined in a plug-and-play differentiable ranking metrics. A pre-trained differen- manner with existing deep architectures. We demonstrate tiable sorter (deep neural net [DNN] ΘB) is used to con- the interest of our approach in three different tasks that vert into ranks the raw scores given by the model (DNN require ranking: Cross-modal text-image retrieval, multi- ΘA) being trained to a collection of inputs. A loss is then label image classification and visual memorability ranking. applied to the predicted rank and the error can be back- Our approach yields very competitive results on these three propagated through the differentiable sorter and used to up- tasks, which validates the merit and the flexibility of SoDeep date the weights ΘA.
    [Show full text]
  • Spectrum-Enhanced Pairwise Learning to Rank
    Spectrum-enhanced Pairwise Learning to Rank Wenhui Yu Zheng Qin∗ Tsinghua University Tsinghua University Beijing, China Beijing, China [email protected] [email protected] ABSTRACT 1 INTRODUCTION To enhance the performance of the recommender system, side in- Recommender systems have been widely used in online services formation is extensively explored with various features (e.g., visual such as E-commerce and social media sites to predict users’ prefer- features and textual features). However, there are some demerits ence based on their interaction histories. Modern recommender sys- of side information: (1) the extra data is not always available in all tems uncover the underlying latent factors that encode the prefer- recommendation tasks; (2) it is only for items, there is seldom high- ence of users and properties of items. In recent years, to strengthen level feature describing users. To address these gaps, we introduce the presentation ability of the models, various features are incor- the spectral features extracted from two hypergraph structures of porated for additional information, such as visual features from the purchase records. Spectral features describe the similarity of product images [14, 44, 48], textual features from review data [8, 9], users/items in the graph space, which is critical for recommendation. and auditory features from music [5]. However, these features are We leverage spectral features to model the users’ preference and not generally applicable, for example, visual features can only be items’ properties by incorporating them into a Matrix Factorization used in product recommendation, while not in music recommenda- (MF) model. tion. Also, these features are unavailable in some recommendation In addition to modeling, we also use spectral features to optimize.
    [Show full text]
  • Embedding Meta-Textual Information for Improved Learning to Rank
    Embedding Meta-Textual Information for Improved Learning to Rank Toshitaka Kuwa∗, Shigehiko Schamoniy;z, Stefan Riezlery;z yDepartment of Computational Linguistics, Heidelberg, Germany. zInterdisciplinary Center for Scientific Computing (IWR), Heidelberg, Germany fkuwa, schamoni, [email protected] Abstract Neural approaches to learning term embeddings have led to improved computation of similarity and ranking in information retrieval (IR). So far neural representation learning has not been ex- tended to meta-textual information that is readily available for many IR tasks, for example, patent classes in prior-art retrieval, topical information in Wikipedia articles, or product categories in e-commerce data. We present a framework that learns embeddings for meta-textual categories, and optimizes a pairwise ranking objective for improved matching based on combined embed- dings of textual and meta-textual information. We show considerable gains in an experimental evaluation on cross-lingual retrieval in the Wikipedia domain for three language pairs, and in the Patent domain for one language pair. Our results emphasize that the mode of combining different types of information is crucial for model improvement. 1 Introduction The recent success of neural methods in IR rests on bridging the gap between query and document vocabulary by learning term embeddings that allow for improved computation of similarity and ranking (Mitra and Craswell, 2018). So far vector representations in neural IR have been confined to textual information, neglecting information beyond the raw text that is potentially useful for retrieval and is readily available in many data situations. For example, meta-textual information is available in the form of patent classes in prior-art search, topic classes in search over Wikipedia articles, or product classes in retrieval in the e-commerce domain.
    [Show full text]
  • Self-Supervised Learning from Permutations Via Differentiable Ranking
    Under review as a conference paper at ICLR 2021 SHUFFLE TO LEARN:SELF-SUPERVISED LEARNING FROM PERMUTATIONS VIA DIFFERENTIABLE RANKING Anonymous authors Paper under double-blind review ABSTRACT Self-supervised pre-training using so-called “pretext” tasks has recently shown impressive performance across a wide range of tasks. In this work, we advance self-supervised learning from permutations, that consists in shuffling parts of in- put and training a model to reorder them, improving downstream performance in classification. To do so, we overcome the main challenges of integrating permu- tation inversions (a discontinuous operation) into an end-to-end training scheme, heretofore sidestepped by casting the reordering task as classification, fundamen- tally reducing the space of permutations that can be exploited. We make two main contributions. First, we use recent advances in differentiable ranking to integrate the permutation inversion flawlessly into a neural network, enabling us to use the full set of permutations, at no additional computing cost. Our experiments validate that learning from all possible permutations improves the quality of the pre-trained representations over using a limited, fixed set. Second, we successfully demon- strate that inverting permutations is a meaningful pretext task in a diverse range of modalities, beyond images, which does not require modality-specific design. In particular, we improve music understanding by reordering spectrogram patches in the time-frequency space, as well as video classification by reordering frames along the time axis. We furthermore analyze the influence of the patches that we use (vertical, horizontal, 2-dimensional), as well as the benefit of our approach in different data regimes.
    [Show full text]